A Deep Learning Framework for Physical-Layer Secure Beamforming
Zihan Song, Yang Lu, Xianhao Chen, Bo Ai, Zhong Zhangdui, Dusit Niyato
Abstract
This paper investigates the deep learning (DL) based physical-layer secure beamforming design. A uniform DL framework is proposed, which exploits training set across various system utilities and enables transfer learning among them. Specifically, a convolutional neural network (CNN) based model named SecCNN and a graph neural network (GNN) based model named SecGNN are respectively designed to map channel vectors to beamforming and artificial noise vectors. The SecCNN adopts circular padding and full-size kernels to capture the global information, and the SecGNN adopts graph partition and semantic attention to distinguish different types of users. The models are trained via unsupervised learning. Numerical results evaluate the models in terms of the optimality, scalability, inference time, stability and transfer learning, which attains superior performance in various settings.